For grid-scale energy storage and national energy resilience, the U.S. needs better batteries. Lawrence Livermore National Laboratory (LLNL) scientists are tackling that challenge in many ways, but one approach is making a significant impact: physics-informed machine learning.
In two recent publications, LLNL researchers examined how integrating molecular dynamics simulations with physics-informed machine learning can illuminate the relationships between structure and behavior in complex battery materials. They used the powerful combination of techniques to explore carbon anodes in sodium-ion batteries and liquid electrolytes in lithium-ion batteries.
"These studies show that the structural complexity of battery materials is not just an obstacle to understanding but a design advantage, laying the groundwork for high-throughput screening of next-generation energy-storage materials," said LLNL scientist and author Liwen (Sabrina) Wan. "By encoding that complexity into physics-informed machine learning models, we can predict properties and identify design levers that traditional approaches simply cannot access."
The first paper, published in Energy Storage Materials, examines sodium-ion batteries. Because sodium is abundant and domestically available, this technology is important for ensuring a robust U.S. supply chain.
Sodium batteries work by moving sodium ions back and forth from anode to cathode. The most commercially mature sodium anodes are made of hard carbon, which looks like a jumble of crumpled, disordered, graphene-like sheets. That structural disorder, full of tiny pores and empty spaces, makes the anode difficult to characterize and engineer.
"Sodium ions can move into all of that disorder, slipping between layers, settling on surfaces and filling nanopores," said LLNL scientist and author Nikhil Rampal. "That complexity is part of what makes hard carbon so promising, but it is also what makes it so challenging to design."
Researchers have long struggled to understand how the atomic features within the hard carbon relate to the transport of sodium ions. In this work, the team used LLNL's high-performance computing to simulate how every atom in the material moves and interacts over time.
"We essentially created an atom-by-atom movie of sodium ions diffusing, clustering or becoming trapped inside the carbon," said Rampal.
Then, the authors used those movies to train a machine learning algorithm to predict how the atoms interact. That algorithm can run much larger, longer and more accurate simulations affordably. It was used to classify sodium ion motion into eight different regimes based on their unique interactions with the hard carbon.
"As carbon density and sodium loading increase, ions cluster or become trapped in nanopores, with direct implications for rate capability and thermal safety," said Rampal.
The result is a quantitative map between microstructure and ion transport that includes actionable ways to enhance the hard carbon. The researchers believe this work provides a concrete path to safely maximize the movement of sodium ions, and therefore the deployment of sodium battery technology.
The second paper, published in EES Batteries, applies the same philosophy to a different challenge: better electrolytes for lithium-ion batteries. Designing an ideal electrolyte is a combinatorial challenge because the endless possibilities of solvents, salts, additives and concentrations are too vast to screen exhaustively.
Conventional electrolyte models rely on text-based representations that neglect the 3D geometry of molecules. In contrast, the LLNL team generated realistic, 3D configurations of molecules with their molecular dynamics simulations. They fed those structures into a machine learning model, which predicted the statistical stability of each configuration.
The key insight is that electrochemical stability depends on the full ensemble of molecules, not just the sum of its parts.
"The salt or solvent identity and concentration can shift the predicted stability window dramatically, through mechanisms that text-based encoders simply cannot see," said Rampal. "For example, swapping one lithium salt for another produced a 57% wider stability window, driven entirely by how the anion arranges itself around the lithium ion."
The scientists envision this molecular dynamics and physics-informed machine learning pipeline as a high-throughput screening platform that replaces trial-and-error electrolyte design with physics-guided exploration. Incorporating experimental benchmark data would sharpen model accuracy over time, and the core principles transfer naturally to other battery chemistries and electrochemical systems.
"For national lab programs exploring large design spaces across lithium, sodium and multivalent battery chemistries, this could significantly accelerate discovery," said Wan. "While these studies focus on batteries, the broader framework can be applied to many other systems."
Other LLNL authors for the hard carbon work include Stephen Weitzner, Marissa Wood and Jonathan RI Lee. Funding support is partially provided by LLNL's Laboratory Directed Research and Development program and partially by the U.S. Department of Energy (DOE), Office of Electricity, Energy Storage Division. Computing support was provided by the LLNL Institutional Computing Grand Challenge program and resources were sponsored by the DOE Office of Critical Minerals and Energy Innovation and located at the National Laboratory of the Rockies.
